Merge branch 'master' into saving-and-resuming

examples/README.md

@@ -43,7 +43,7 @@ Quick benchmarks from the script (no other modifications):
 | Titan V | AMP | 26s | 0.8281/0.8568/0.8411 |
 | V100    | FP32 | 35s | 0.8646/0.8359/0.8464 |
 | V100    | AMP | 22s | 0.8646/0.8385/0.8411 |
-| 1080 Ti | FP32 | 55s | - | 
+| 1080 Ti | FP32 | 55s | - |
 
 Mixed precision (AMP) reduces the training time considerably for the same hardware and hyper-parameters (same batch size was used).
 
@@ -357,9 +357,9 @@ eval_loss = 0.44457291918821606
 
 Based on the script [`run_squad.py`](https://github.com/huggingface/transformers/blob/master/examples/run_squad.py).
 
-#### Fine-tuning on SQuAD
+#### Fine-tuning BERT on SQuAD1.0
 
-This example code fine-tunes BERT on the SQuAD dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) 
+This example code fine-tunes BERT on the SQuAD1.0 dataset. It runs in 24 min (with BERT-base) or 68 min (with BERT-large) 
 on a single tesla V100 16GB. The data for SQuAD can be downloaded with the following links and should be saved in a 
 $SQUAD_DIR directory.
 
@@ -367,6 +367,12 @@ $SQUAD_DIR directory.
 * [dev-v1.1.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v1.1.json)
 * [evaluate-v1.1.py](https://github.com/allenai/bi-att-flow/blob/master/squad/evaluate-v1.1.py)
 
+And for SQuAD2.0, you need to download:
+
+- [train-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/train-v2.0.json)
+- [dev-v2.0.json](https://rajpurkar.github.io/SQuAD-explorer/dataset/dev-v2.0.json)
+- [evaluate-v2.0.py](https://worksheets.codalab.org/rest/bundles/0x6b567e1cf2e041ec80d7098f031c5c9e/contents/blob/)
+
 ```bash
 export SQUAD_DIR=/path/to/SQUAD
 
@@ -396,7 +402,7 @@ exact_match = 81.22
 #### Distributed training
 
 
-Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD:
+Here is an example using distributed training on 8 V100 GPUs and Bert Whole Word Masking uncased model to reach a F1 > 93 on SQuAD1.0:
 
 ```bash
 python -m torch.distributed.launch --nproc_per_node=8 run_squad.py \
@@ -428,7 +434,9 @@ This fine-tuned model is available as a checkpoint under the reference
 
 #### Fine-tuning XLNet on SQuAD
 
-This example code fine-tunes XLNet on the SQuAD dataset. See above to download the data for SQuAD .
+This example code fine-tunes XLNet on both SQuAD1.0 and SQuAD2.0 dataset. See above to download the data for SQuAD .
+
+##### Command for SQuAD1.0:
 
 ```bash
 export SQUAD_DIR=/path/to/SQUAD
@@ -451,7 +459,32 @@ python /data/home/hlu/transformers/examples/run_squad.py \
     --save_steps 5000
 ```
 
-Training with the previously defined hyper-parameters yields the following results:
+##### Command for SQuAD2.0:
+
+```bash
+export SQUAD_DIR=/path/to/SQUAD
+
+python run_squad.py \
+    --model_type xlnet \
+    --model_name_or_path xlnet-large-cased \
+    --do_train \
+    --do_eval \
+    --version_2_with_negative \
+    --train_file $SQUAD_DIR/train-v2.0.json \
+    --predict_file $SQUAD_DIR/dev-v2.0.json \
+    --learning_rate 3e-5 \
+    --num_train_epochs 4 \
+    --max_seq_length 384 \
+    --doc_stride 128 \
+    --output_dir ./wwm_cased_finetuned_squad/ \
+    --per_gpu_eval_batch_size=2  \
+    --per_gpu_train_batch_size=2   \
+    --save_steps 5000
+```
+
+Larger batch size may improve the performance while costing more memory.
+
+##### Results for SQuAD1.0 with the previously defined hyper-parameters:
 
 ```python
 {
@@ -464,10 +497,28 @@ Training with the previously defined hyper-parameters yields the following resul
 }
 ```
 
+##### Results for SQuAD2.0 with the previously defined hyper-parameters:
+
+```python
+{
+"exact": 80.4177545691906,
+"f1": 84.07154997729623,
+"total": 11873,
+"HasAns_exact": 76.73751686909581,
+"HasAns_f1": 84.05558584352873,
+"HasAns_total": 5928,
+"NoAns_exact": 84.0874684608915,
+"NoAns_f1": 84.0874684608915,
+"NoAns_total": 5945
+}
+```
+
+
+
 ## Named Entity Recognition
 
 Based on the scripts [`run_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_ner.py) for Pytorch and
-[`run_tf_ner.py`(https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py)] for Tensorflow 2.
+[`run_tf_ner.py`](https://github.com/huggingface/transformers/blob/master/examples/run_tf_ner.py) for Tensorflow 2.
 This example fine-tune Bert Multilingual on GermEval 2014 (German NER).
 Details and results for the fine-tuning provided by @stefan-it.
 

examples/run_generation.py

@@ -20,14 +20,10 @@ from __future__ import absolute_import, division, print_function, unicode_litera
 
 import argparse
 import logging
-from tqdm import trange
 
 import torch
-import torch.nn.functional as F
 import numpy as np
 
-from transformers import GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig
-
 from transformers import GPT2LMHeadModel, GPT2Tokenizer
 from transformers import OpenAIGPTLMHeadModel, OpenAIGPTTokenizer
 from transformers import XLNetLMHeadModel, XLNetTokenizer
@@ -36,22 +32,22 @@ from transformers import CTRLLMHeadModel, CTRLTokenizer
 from transformers import XLMWithLMHeadModel, XLMTokenizer
 
 
-logging.basicConfig(format = '%(asctime)s - %(levelname)s - %(name)s -   %(message)s',
-                    datefmt = '%m/%d/%Y %H:%M:%S',
-                    level = logging.INFO)
+logging.basicConfig(
+    format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+    datefmt="%m/%d/%Y %H:%M:%S",
+    level=logging.INFO,
+)
 logger = logging.getLogger(__name__)
 
 MAX_LENGTH = int(10000)  # Hardcoded max length to avoid infinite loop
 
-ALL_MODELS = sum((tuple(conf.pretrained_config_archive_map.keys()) for conf in (GPT2Config, OpenAIGPTConfig, XLNetConfig, TransfoXLConfig, XLMConfig, CTRLConfig)), ())
-
 MODEL_CLASSES = {
-    'gpt2': (GPT2LMHeadModel, GPT2Tokenizer),
-    'ctrl': (CTRLLMHeadModel, CTRLTokenizer),
-    'openai-gpt': (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
-    'xlnet': (XLNetLMHeadModel, XLNetTokenizer),
-    'transfo-xl': (TransfoXLLMHeadModel, TransfoXLTokenizer),
-    'xlm': (XLMWithLMHeadModel, XLMTokenizer),
+    "gpt2": (GPT2LMHeadModel, GPT2Tokenizer),
+    "ctrl": (CTRLLMHeadModel, CTRLTokenizer),
+    "openai-gpt": (OpenAIGPTLMHeadModel, OpenAIGPTTokenizer),
+    "xlnet": (XLNetLMHeadModel, XLNetTokenizer),
+    "transfo-xl": (TransfoXLLMHeadModel, TransfoXLTokenizer),
+    "xlm": (XLMWithLMHeadModel, XLMTokenizer),
 }
 
 # Padding text to help Transformer-XL and XLNet with short prompts as proposed by Aman Rusia
@@ -75,81 +71,79 @@ def set_seed(args):
     if args.n_gpu > 0:
         torch.cuda.manual_seed_all(args.seed)
 
-
-def top_k_top_p_filtering(logits, top_k=0, top_p=0.0, filter_value=-float('Inf')):
-    """ Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
-        Args:
-            logits: logits distribution shape (batch size x vocabulary size)
-            top_k > 0: keep only top k tokens with highest probability (top-k filtering).
-            top_p > 0.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
-                Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
-        From: https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
-    """
-    top_k = min(top_k, logits.size(-1))  # Safety check
-    if top_k > 0:
-        # Remove all tokens with a probability less than the last token of the top-k
-        indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
-        logits[indices_to_remove] = filter_value
-
-    if top_p > 0.0:
-        sorted_logits, sorted_indices = torch.sort(logits, descending=True)
-        cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
-
-        # Remove tokens with cumulative probability above the threshold
-        sorted_indices_to_remove = cumulative_probs > top_p
-        # Shift the indices to the right to keep also the first token above the threshold
-        sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
-        sorted_indices_to_remove[..., 0] = 0
-
-        # scatter sorted tensors to original indexing
-        indices_to_remove = sorted_indices_to_remove.scatter(dim=1, index=sorted_indices, src=sorted_indices_to_remove)
-        logits[indices_to_remove] = filter_value
-    return logits
+#
+# Functions to prepare models' input
+#
 
 
-def sample_sequence(model, length, context, num_samples=1, temperature=1, top_k=0, top_p=0.0, repetition_penalty=1.0,
-                    is_xlnet=False, is_xlm_mlm=False, xlm_mask_token=None, xlm_lang=None, device='cpu'):
-    context = torch.tensor(context, dtype=torch.long, device=device)
-    context = context.unsqueeze(0).repeat(num_samples, 1)
-    generated = context
-    with torch.no_grad():
-        for _ in trange(length):
+def prepare_ctrl_input(args, _, tokenizer, prompt_text):
+    if args.temperature > 0.7:
+        logger.info(
+            "CTRL typically works better with lower temperatures (and lower top_k)."
+        )
 
-            inputs = {'input_ids': generated}
-            if is_xlnet: 
-                # XLNet is a direct (predict same token, not next token) and bi-directional model by default
-                # => need one additional dummy token in the input (will be masked), attention mask and target mapping (see model docstring)
-                input_ids = torch.cat((generated, torch.zeros((1, 1), dtype=torch.long, device=device)), dim=1)
-                perm_mask = torch.zeros((1, input_ids.shape[1], input_ids.shape[1]), dtype=torch.float, device=device)
-                perm_mask[:, :, -1] = 1.0  # Previous tokens don't see last token
-                target_mapping = torch.zeros((1, 1, input_ids.shape[1]), dtype=torch.float, device=device)
-                target_mapping[0, 0, -1] = 1.0  # predict last token
-                inputs = {'input_ids': input_ids, 'perm_mask': perm_mask, 'target_mapping': target_mapping}
+    encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False)
+    if not any(encoded_prompt[0] == x for x in tokenizer.control_codes.values()):
+        logger.info(
+            "WARNING! You are not starting your generation from a control code so you won't get good results"
+        )
+    return prompt_text
 
-            if is_xlm_mlm and xlm_mask_token:
-                # XLM MLM models are direct models (predict same token, not next token)
-                # => need one additional dummy token in the input (will be masked and guessed)
-                input_ids = torch.cat((generated, torch.full((1, 1), xlm_mask_token, dtype=torch.long, device=device)), dim=1)
-                inputs = {'input_ids': input_ids}
 
-            if xlm_lang is not None:
-                inputs["langs"] = torch.tensor([xlm_lang] * inputs["input_ids"].shape[1], device=device).view(1, -1)
+def prepare_xlm_input(args, model, tokenizer, prompt_text):
+    # kwargs = {"language": None, "mask_token_id": None}
 
-            outputs = model(**inputs)  # Note: we could also use 'past' with GPT-2/Transfo-XL/XLNet/CTRL (cached hidden-states)
-            next_token_logits = outputs[0][:, -1, :] / (temperature if temperature > 0 else 1.)
+    # Set the language
+    use_lang_emb = hasattr(model.config, "use_lang_emb") and model.config.use_lang_emb
+    if hasattr(model.config, "lang2id") and use_lang_emb:
+        available_languages = model.config.lang2id.keys()
+        if args.xlm_language in available_languages:
+            language = args.xlm_language
+        else:
+            language = None
+            while language not in available_languages:
+                language = input(
+                    "Using XLM. Select language in "
+                    + str(list(available_languages))
+                    + " >>> "
+                )
+        # kwargs["language"] = tokenizer.lang2id[language]
 
-            # repetition penalty from CTRL (https://arxiv.org/abs/1909.05858)
-            for i in range(num_samples):
-                for _ in set(generated[i].tolist()):
-                    next_token_logits[i, _] /= repetition_penalty
-                
-            filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=top_k, top_p=top_p)
-            if temperature == 0: # greedy sampling:
-                next_token = torch.argmax(filtered_logits, dim=-1).unsqueeze(-1)
-            else:
-                next_token = torch.multinomial(F.softmax(filtered_logits, dim=-1), num_samples=1)
-            generated = torch.cat((generated, next_token), dim=1)
-    return generated
+    # TODO fix mask_token_id setup when configurations will be synchronized between models and tokenizers
+    # XLM masked-language modeling (MLM) models need masked token
+    # is_xlm_mlm = "mlm" in args.model_name_or_path
+    # if is_xlm_mlm:
+    #     kwargs["mask_token_id"] = tokenizer.mask_token_id
+
+    return prompt_text
+
+
+def prepare_xlnet_input(args, _, tokenizer, prompt_text):
+    prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
+    return prompt_text, {}
+
+
+def prepare_transfoxl_input(args, _, tokenizer, prompt_text):
+    prompt_text = (args.padding_text if args.padding_text else PADDING_TEXT) + prompt_text
+    return prompt_text, {}
+
+
+PREPROCESSING_FUNCTIONS = {
+    "ctrl": prepare_ctrl_input,
+    "xlm": prepare_xlm_input,
+    "xlnet": prepare_xlnet_input,
+    "transfo-xl": prepare_transfoxl_input,
+}
+
+
+def adjust_length_to_model(length, max_sequence_length):
+    if length < 0 and max_sequence_length > 0:
+        length = max_sequence_length
+    elif 0 < max_sequence_length < length:
+        length = max_sequence_length  # No generation bigger than model size
+    elif length < 0:
+        length = MAX_LENGTH  # avoid infinite loop
+    return length
 
 
 def main():
@@ -157,104 +151,76 @@ def main():
     parser.add_argument("--model_type", default=None, type=str, required=True,
                         help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
     parser.add_argument("--model_name_or_path", default=None, type=str, required=True,
-                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(ALL_MODELS))
+                        help="Path to pre-trained model or shortcut name selected in the list: " + ", ".join(MODEL_CLASSES.keys()))
+
     parser.add_argument("--prompt", type=str, default="")
-    parser.add_argument("--padding_text", type=str, default="")
-    parser.add_argument("--xlm_lang", type=str, default="", help="Optional language when used with the XLM model.")
     parser.add_argument("--length", type=int, default=20)
-    parser.add_argument("--num_samples", type=int, default=1)
-    parser.add_argument("--temperature", type=float, default=1.0,
-                        help="temperature of 0 implies greedy sampling")
-    parser.add_argument("--repetition_penalty", type=float, default=1.0,
-                        help="primarily useful for CTRL model; in that case, use 1.2")
-    parser.add_argument("--top_k", type=int, default=0)
-    parser.add_argument("--top_p", type=float, default=0.9)
-    parser.add_argument("--no_cuda", action='store_true',
-                        help="Avoid using CUDA when available")
-    parser.add_argument('--seed', type=int, default=42,
-                        help="random seed for initialization")
-    parser.add_argument('--stop_token', type=str, default=None,
-                        help="Token at which text generation is stopped")
+    parser.add_argument("--stop_token", type=str, default=None, help="Token at which text generation is stopped")
+
+    parser.add_argument("--temperature", type=float, default=1.0, help="temperature of 1.0 has no effect, lower tend toward greedy sampling")
+    parser.add_argument("--repetition_penalty", type=float, default=1.0, help="primarily useful for CTRL model; in that case, use 1.2")
+    parser.add_argument("--k", type=int, default=0)
+    parser.add_argument("--p", type=float, default=0.9)
+
+    parser.add_argument("--padding_text", type=str, default="", help="Padding text for Transfo-XL and XLNet.")
+    parser.add_argument("--xlm_language", type=str, default="", help="Optional language when used with the XLM model.")
+
+    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+    parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
     args = parser.parse_args()
 
-    args.device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+    args.device = torch.device(
+        "cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu"
+    )
     args.n_gpu = torch.cuda.device_count()
 
     set_seed(args)
 
-    args.model_type = args.model_type.lower()
-    model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    # Initialize the model and tokenizer
+    try:
+        args.model_type = args.model_type.lower()
+        model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    except KeyError:
+        raise KeyError(
+            "the model {} you specified is not supported. You are welcome to add it and open a PR :)"
+        )
+
     tokenizer = tokenizer_class.from_pretrained(args.model_name_or_path)
     model = model_class.from_pretrained(args.model_name_or_path)
     model.to(args.device)
-    model.eval()
-
-    if args.length < 0 and model.config.max_position_embeddings > 0:
-        args.length = model.config.max_position_embeddings
-    elif 0 < model.config.max_position_embeddings < args.length:
-        args.length = model.config.max_position_embeddings  # No generation bigger than model size 
-    elif args.length < 0:
-        args.length = MAX_LENGTH  # avoid infinite loop
 
+    args.length = adjust_length_to_model(
+        args.length, max_sequence_length=model.config.max_position_embeddings
+    )
     logger.info(args)
-    if args.model_type in ["ctrl"]:
-        if args.temperature > 0.7:
-            logger.info('CTRL typically works better with lower temperatures (and lower top_k).')
 
-    while True:
-        xlm_lang = None
-        # XLM Language usage detailed in the issues #1414
-        if args.model_type in ["xlm"] and hasattr(tokenizer, 'lang2id') and hasattr(model.config, 'use_lang_emb') \
-                and model.config.use_lang_emb:
-            if args.xlm_lang:
-                language = args.xlm_lang
-            else:
-                language = None
-                while language not in tokenizer.lang2id.keys():
-                    language = input("Using XLM. Select language in " + str(list(tokenizer.lang2id.keys())) + " >>> ")
-            xlm_lang = tokenizer.lang2id[language]
+    prompt_text = args.prompt if args.prompt else input("Model prompt >>> ")
 
-        # XLM masked-language modeling (MLM) models need masked token (see details in sample_sequence)
-        is_xlm_mlm = args.model_type in ["xlm"] and 'mlm' in args.model_name_or_path
-        if is_xlm_mlm:
-            xlm_mask_token = tokenizer.mask_token_id
-        else:
-            xlm_mask_token = None
+    # Different models need different input formatting and/or extra arguments
+    requires_preprocessing = args.model_type in PREPROCESSING_FUNCTIONS.keys()
+    if requires_preprocessing:
+        prepare_input = PREPROCESSING_FUNCTIONS.get(args.model_type)
+        prompt_text = prepare_input(args, model, tokenizer, prompt_text)
+    encoded_prompt = tokenizer.encode(prompt_text, add_special_tokens=False, return_tensors='pt')
 
-        raw_text = args.prompt if args.prompt else input("Model prompt >>> ")
-        if args.model_type in ["transfo-xl", "xlnet"]:
-            # Models with memory likes to have a long prompt for short inputs.
-            raw_text = (args.padding_text if args.padding_text else PADDING_TEXT) + raw_text
-        context_tokens = tokenizer.encode(raw_text, add_special_tokens=False)
-        if args.model_type == "ctrl":
-            if not any(context_tokens[0] == x for x in tokenizer.control_codes.values()):
-                logger.info("WARNING! You are not starting your generation from a control code so you won't get good results")
-        out = sample_sequence(
-            model=model,
-            context=context_tokens,
-            num_samples=args.num_samples,
-            length=args.length,
-            temperature=args.temperature,
-            top_k=args.top_k,
-            top_p=args.top_p,
-            repetition_penalty=args.repetition_penalty,
-            is_xlnet=bool(args.model_type == "xlnet"),
-            is_xlm_mlm=is_xlm_mlm,
-            xlm_mask_token=xlm_mask_token,
-            xlm_lang=xlm_lang,
-            device=args.device,
-        )
-        out = out[:, len(context_tokens):].tolist()
-        for o in out:
-            text = tokenizer.decode(o, clean_up_tokenization_spaces=True)
-            text = text[: text.find(args.stop_token) if args.stop_token else None]
+    output_sequences = model.generate(
+        input_ids=encoded_prompt,
+        max_length=args.length,
+        temperature=args.temperature,
+        top_k=args.k,
+        top_p=args.p,
+        repetition_penalty=args.repetition_penalty,
+    )
 
-            print(text)
+    # Batch size == 1. to add more examples please use num_return_sequences > 1 
+    generated_sequence = output_sequences[0].tolist()
+    text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True)
+    text = text[: t.find(args.stop_token) if args.stop_token else None]
+
+    print(text)
 
-        if args.prompt:
-            break
     return text
 
 
-if __name__ == '__main__':
+if __name__ == "__main__":
     main()

examples/run_glue.py

@@ -52,6 +52,9 @@ from transformers import (WEIGHTS_NAME, BertConfig,
                                   AlbertConfig,
                                   AlbertForSequenceClassification, 
                                   AlbertTokenizer,
+                                  XLMRobertaConfig,
+                                  XLMRobertaForSequenceClassification,
+                                  XLMRobertaTokenizer,
                                 )
 
 from transformers import AdamW, get_linear_schedule_with_warmup
@@ -72,7 +75,8 @@ MODEL_CLASSES = {
     'xlm': (XLMConfig, XLMForSequenceClassification, XLMTokenizer),
     'roberta': (RobertaConfig, RobertaForSequenceClassification, RobertaTokenizer),
     'distilbert': (DistilBertConfig, DistilBertForSequenceClassification, DistilBertTokenizer),
-    'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer)
+    'albert': (AlbertConfig, AlbertForSequenceClassification, AlbertTokenizer),
+    'xlmroberta': (XLMRobertaConfig, XLMRobertaForSequenceClassification, XLMRobertaTokenizer),
 }
 
 
@@ -337,9 +341,9 @@ def load_and_cache_examples(args, task, tokenizer, evaluate=False):
     else:
         logger.info("Creating features from dataset file at %s", args.data_dir)
         label_list = processor.get_labels()
-        if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta']:
+        if task in ['mnli', 'mnli-mm'] and args.model_type in ['roberta', 'xlmroberta']:
             # HACK(label indices are swapped in RoBERTa pretrained model)
-            label_list[1], label_list[2] = label_list[2], label_list[1] 
+            label_list[1], label_list[2] = label_list[2], label_list[1]
         examples = processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
         features = convert_examples_to_features(examples,
                                                 tokenizer,
@@ -413,7 +417,7 @@ def main():
     parser.add_argument("--learning_rate", default=5e-5, type=float,
                         help="The initial learning rate for Adam.")
     parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight deay if we apply some.")
+                        help="Weight decay if we apply some.")
     parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                         help="Epsilon for Adam optimizer.")
     parser.add_argument("--max_grad_norm", default=1.0, type=float,

examples/run_lm_finetuning.py

@@ -430,7 +430,7 @@ def main():
     parser.add_argument("--learning_rate", default=5e-5, type=float,
                         help="The initial learning rate for Adam.")
     parser.add_argument("--weight_decay", default=0.0, type=float,
-                        help="Weight deay if we apply some.")
+                        help="Weight decay if we apply some.")
     parser.add_argument("--adam_epsilon", default=1e-8, type=float,
                         help="Epsilon for Adam optimizer.")
     parser.add_argument("--max_grad_norm", default=1.0, type=float,

examples/run_ner.py

@@ -38,11 +38,13 @@ from transformers import WEIGHTS_NAME, BertConfig, BertForTokenClassification, B
 from transformers import RobertaConfig, RobertaForTokenClassification, RobertaTokenizer
 from transformers import DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer
 from transformers import CamembertConfig, CamembertForTokenClassification, CamembertTokenizer
+from transformers import XLMRobertaConfig, XLMRobertaForTokenClassification, XLMRobertaTokenizer
 
 logger = logging.getLogger(__name__)
 
 ALL_MODELS = sum(
-    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig)),
+    (tuple(conf.pretrained_config_archive_map.keys()) for conf in (BertConfig, RobertaConfig, DistilBertConfig,
+                                                                   CamembertConfig, XLMRobertaConfig)),
     ())
 
 MODEL_CLASSES = {
@@ -50,6 +52,7 @@ MODEL_CLASSES = {
     "roberta": (RobertaConfig, RobertaForTokenClassification, RobertaTokenizer),
     "distilbert": (DistilBertConfig, DistilBertForTokenClassification, DistilBertTokenizer),
     "camembert": (CamembertConfig, CamembertForTokenClassification, CamembertTokenizer),
+    "xlmroberta": (XLMRobertaConfig, XLMRobertaForTokenClassification, XLMRobertaTokenizer),
 }
 
 

examples/run_squad.py

@@ -62,7 +62,6 @@ MODEL_CLASSES = {
     'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer),
     'distilbert': (DistilBertConfig, DistilBertForQuestionAnswering, DistilBertTokenizer),
     'albert': (AlbertConfig, AlbertForQuestionAnswering, AlbertTokenizer),
-    'xlm': (XLMConfig, XLMForQuestionAnswering, XLMTokenizer)
 }
 
 
@@ -200,8 +199,10 @@ def train(args, train_dataset, model, tokenizer):
                 inputs['token_type_ids'] = None if args.model_type == 'xlm' else batch[2]
 
             if args.model_type in ['xlnet', 'xlm']:
-                inputs.update({'cls_index': batch[5], 'p_mask': batch[6]})
-
+                inputs.update({'cls_index': batch[5],
+                               'p_mask':       batch[6]})
+                if args.version_2_with_negative:
+                    inputs.update({'is_impossible': batch[7]})
             outputs = model(**inputs)
             # model outputs are always tuple in transformers (see doc)
             loss = outputs[0]
@@ -296,7 +297,7 @@ def evaluate(args, model, tokenizer, prefix=""):
         dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
 
     # multi-gpu evaluate
-    if args.n_gpu > 1:
+    if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
         model = torch.nn.DataParallel(model)
 
     # Eval!
@@ -420,7 +421,7 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
     else:
         logger.info("Creating features from dataset file at %s", input_dir)
 
-        if not args.data_dir:
+        if not args.data_dir and ((evaluate and not args.predict_file) or (not evaluate and not args.train_file)):
             try:
                 import tensorflow_datasets as tfds
             except ImportError:
@@ -436,8 +437,10 @@ def load_and_cache_examples(args, tokenizer, evaluate=False, output_examples=Fal
                 tfds_examples, evaluate=evaluate)
         else:
             processor = SquadV2Processor() if args.version_2_with_negative else SquadV1Processor()
-            examples = processor.get_dev_examples(
-                args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
+            if evaluate:
+                examples = processor.get_dev_examples(args.data_dir, filename=args.predict_file)
+            else:
+                examples = processor.get_train_examples(args.data_dir, filename=args.train_file)
 
         features, dataset = squad_convert_examples_to_features(
             examples=examples,
@@ -477,7 +480,14 @@ def main():
 
     # Other parameters
     parser.add_argument("--data_dir", default=None, type=str,
-                        help="The input data dir. Should contain the .json files for the task. If not specified, will run with tensorflow_datasets.")
+                        help="The input data dir. Should contain the .json files for the task." +
+                             "If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
+    parser.add_argument("--train_file", default=None, type=str,
+                        help="The input training file. If a data dir is specified, will look for the file there" +
+                             "If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
+    parser.add_argument("--predict_file", default=None, type=str,
+                        help="The input evaluation file. If a data dir is specified, will look for the file there" +
+                             "If no data dir or train/predict files are specified, will run with tensorflow_datasets.")
     parser.add_argument("--config_name", default="", type=str,
                         help="Pretrained config name or path if not the same as model_name")
     parser.add_argument("--tokenizer_name", default="", type=str,
@@ -564,11 +574,6 @@ def main():
                         help="Can be used for distant debugging.")
     args = parser.parse_args()
 
-    args.predict_file = os.path.join(args.output_dir, 'predictions_{}_{}.txt'.format(
-        list(filter(None, args.model_name_or_path.split('/'))).pop(),
-        str(args.max_seq_length))
-    )
-
     if os.path.exists(args.output_dir) and os.listdir(args.output_dir) and args.do_train and not args.overwrite_output_dir:
         raise ValueError(
             "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(args.output_dir))
@@ -676,12 +681,15 @@ def main():
     # Evaluation - we can ask to evaluate all the checkpoints (sub-directories) in a directory
     results = {}
     if args.do_eval and args.local_rank in [-1, 0]:
-        checkpoints = [args.output_dir]
-        if args.eval_all_checkpoints:
-            checkpoints = list(os.path.dirname(c) for c in sorted(
-                glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
-            logging.getLogger("transformers.modeling_utils").setLevel(
-                logging.WARN)  # Reduce model loading logs
+        if args.do_train:
+            logger.info("Loading checkpoints saved during training for evaluation")
+            checkpoints = [args.output_dir]
+            if args.eval_all_checkpoints:
+                checkpoints = list(os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + '/**/' + WEIGHTS_NAME, recursive=True)))
+                logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce model loading logs
+        else:
+            logger.info("Loading checkpoint %s for evaluation", args.model_name_or_path)
+            checkpoints = [args.model_name_or_path]
 
         logger.info("Evaluate the following checkpoints: %s", checkpoints)
 

examples/summarization/README.md

@@ -29,7 +29,7 @@ And move all the stories to the same folder. We will refer as `$DATA_PATH` the p
 python run_summarization.py \
     --documents_dir $DATA_PATH \
     --summaries_output_dir $SUMMARIES_PATH \ # optional
-    --to_cpu false \
+    --no_cuda false \
     --batch_size 4 \
     --min_length 50 \
     --max_length 200 \
@@ -39,7 +39,7 @@ python run_summarization.py \
     --compute_rouge true
 ```
 
-The scripts executes on GPU if one is available and if `to_cpu` is not set to `true`. Inference on multiple GPUs is not suported yet. The ROUGE scores will be displayed in the console at the end of evaluation and written in a `rouge_scores.txt` file. The script takes 30 hours to compute with a single Tesla V100 GPU and a batch size of 10 (300,000 texts to summarize).
+The scripts executes on GPU if one is available and if `no_cuda` is not set to `true`. Inference on multiple GPUs is not suported yet. The ROUGE scores will be displayed in the console at the end of evaluation and written in a `rouge_scores.txt` file. The script takes 30 hours to compute with a single Tesla V100 GPU and a batch size of 10 (300,000 texts to summarize).
 
 ## Summarize any text
 
@@ -49,7 +49,7 @@ Put the documents that you would like to summarize in a folder (the path to whic
 python run_summarization.py \
     --documents_dir $DATA_PATH \
     --summaries_output_dir $SUMMARIES_PATH \ # optional
-    --to_cpu false \
+    --no_cuda false \
     --batch_size 4 \
     --min_length 50 \
     --max_length 200 \

examples/summarization/configuration_bertabs.py

@@ -33,6 +33,8 @@ class BertAbsConfig(PretrainedConfig):
     r""" Class to store the configuration of the BertAbs model.
 
     Arguments:
+        vocab_size: int
+            Number of tokens in the vocabulary.
         max_pos: int
             The maximum sequence length that this model will be used with.
         enc_layer: int
@@ -65,7 +67,7 @@ class BertAbsConfig(PretrainedConfig):
 
     def __init__(
         self,
-        vocab_size_or_config_json_file=30522,
+        vocab_size=30522,
         max_pos=512,
         enc_layers=6,
         enc_hidden_size=512,
@@ -81,39 +83,17 @@ class BertAbsConfig(PretrainedConfig):
     ):
         super(BertAbsConfig, self).__init__(**kwargs)
 
-        if self._input_is_path_to_json(vocab_size_or_config_json_file):
-            path_to_json = vocab_size_or_config_json_file
-            with open(path_to_json, "r", encoding="utf-8") as reader:
-                json_config = json.loads(reader.read())
-            for key, value in json_config.items():
-                self.__dict__[key] = value
-        elif isinstance(vocab_size_or_config_json_file, int):
-            self.vocab_size = vocab_size_or_config_json_file
-            self.max_pos = max_pos
+        self.vocab_size = vocab_size
+        self.max_pos = max_pos
 
-            self.enc_layers = enc_layers
-            self.enc_hidden_size = enc_hidden_size
-            self.enc_heads = enc_heads
-            self.enc_ff_size = enc_ff_size
-            self.enc_dropout = enc_dropout
+        self.enc_layers = enc_layers
+        self.enc_hidden_size = enc_hidden_size
+        self.enc_heads = enc_heads
+        self.enc_ff_size = enc_ff_size
+        self.enc_dropout = enc_dropout
 
-            self.dec_layers = dec_layers
-            self.dec_hidden_size = dec_hidden_size
-            self.dec_heads = dec_heads
-            self.dec_ff_size = dec_ff_size
-            self.dec_dropout = dec_dropout
-        else:
-            raise ValueError(
-                "First argument must be either a vocabulary size (int)"
-                "or the path to a pretrained model config file (str)"
-            )
-
-    def _input_is_path_to_json(self, first_argument):
-        """ Checks whether the first argument passed to config
-        is the path to a JSON file that contains the config.
-        """
-        is_python_2 = sys.version_info[0] == 2
-        if is_python_2:
-            return isinstance(first_argument, unicode)
-        else:
-            return isinstance(first_argument, str)
+        self.dec_layers = dec_layers
+        self.dec_hidden_size = dec_hidden_size
+        self.dec_heads = dec_heads
+        self.dec_ff_size = dec_ff_size
+        self.dec_dropout = dec_dropout